Two schools can hand you the same AI policy and be led in opposite directions. Across the leadership teams I work with, the difference never shows up in the document. It shows up when you ask one question: what has this team actually decided about AI since the start of the year, and what changed the week after you decided it? One team answers in about fifteen seconds. The other describes its governance structure.
That gap matters more this term than last. In England, Keeping Children Safe in Education 2026 came into force on September 1, 2026. It brings images that are entirely AI-generated or digitally altered within its guidance on nudes and semi-nudes, and it expects filtering and monitoring to be reviewed at least once every academic year by the senior leader responsible, supported by the designated safeguarding lead and IT support (SWGfL, July 9, 2026). In the United States, districts opened the school year with state model policies already in hand. Nearly everyone now has the documents. Very few can show you the decisions.
What does good AI leadership actually look like?
Good AI leadership looks like five things a visitor could observe in a morning: a short trail of dated decisions, a leader using AI on their own work where staff can see it, something that has visibly stopped, a line the organization will not cross that the leader can state unprompted, and one open question the leader admits to.
None of the five requires a leader to be technical. None can be bought in, delegated to a working group, or copied from another school. That is why they are worth watching: they are the parts of AI leadership that only the leader can supply.
Why an AI strategy document tells you nothing
A document records what a leadership team wanted to be true, which is a different thing from what happened. The sharpest evidence for the gap comes from the people the documents are about. In a survey of 3,020 learners run by AudienceNet for Jobs for the Future, fielded November 28 to December 8, 2025 and published March 18, 2026, only 23 percent named their school or training program as a primary source of information about AI. Forty-eight percent named social media. Thirteen percent did not know what their own institution's policy allowed.
Read those numbers as a measure of leadership rather than of students. The policy exists. The signal did not arrive. When the people a policy is written about cannot say what it says, the document has been produced and the decision has not been led.
The Decision Trail: what to ask for instead of a strategy
The Decision Trail is what I ask a leadership team to show me instead of their AI strategy: the last five decisions they have made about AI, when each was made, who made it, and what changed the week after. Good AI leadership leaves a trail of decisions. Weak AI leadership leaves a document.
This is my suggested way of looking at it rather than a validated instrument. Its value is that it is quick, it can be run in a corridor, and it is close to impossible to fake. A team that is leading fills in five rows from memory and then argues about one of the dates. A team that is not leading fills the rows with things that were announced rather than decided: a working group formed, a webinar attended, a policy adopted, a supplier met. An announcement has no week after. That is how you tell the two apart in under a minute.
Copy this: the Decision Trail template
Copy the layout below into your next leadership meeting and fill in five rows from memory, without opening any document. The two filled rows are hypothetical, there to show the grain a real row should have.
| # | What was decided | Date | Who decided | What changed the week after |
|---|---|---|---|---|
| 1 (hypothetical) | Staff may not paste pupil names into any AI tool except the two approved ones | September 8 | Deputy head | A letter template moved onto an approved tool; one teacher reported a free tool she had been using |
| 2 (hypothetical) | The separate half-termly AI update to governors is folded into the existing headteacher's report | September 22 | Headteacher | One report left the calendar; the governors' agenda item shortened |
| 3 | ||||
| 4 | ||||
| 5 |
Then read it back with three rules:
- A row counts only if it has a date, a name and a week after. If any of the three is missing, it is an intention.
- Announcements do not count. A working group formed, a webinar attended, a policy adopted or a supplier met has no week after, so it cannot go in the table.
- Fewer than three rows from memory is the finding. You do not need a better strategy yet. You need to make one dated decision this fortnight.
The table also makes a short, honest item for the next board paper, because it shows what to report to governors about AI as decisions rather than activity.
The five things you would see
In an organization where AI is being led rather than described, five things are visible to anyone who spends a morning there, and each of them belongs to the leader rather than to the policy.
1. Decisions with dates and names attached
Strong AI leadership produces decisions small enough to date. "We will not use AI to write the first draft of a safeguarding referral, from October, and the deputy head owns it" is a decision. "We are committed to the ethical use of AI" is a value, and values have no week after. Weak leadership talks in commitments because commitments cannot be audited. If your team has made no dated AI decision this term, the honest position is that AI is being managed rather than led.
2. The leader's own work, done where people can see it
Watch what the leader does with AI on their own tasks, in front of others, including the parts it gets wrong. Staff read a leader's private enthusiasm as permission for nobody. They read a governor's report drafted in the open, with the two paragraphs the model got badly wrong shown on the screen, as permission for everyone. That is also the fastest available demonstration that outputs are drafts requiring judgment, which is the thing every policy asserts and almost no policy teaches.
3. Something that stopped
Every real AI decision has a subtraction attached to it. If AI has saved forty minutes a week on reports and nobody has said what those forty minutes are now for, the saving is a rounding error, and staff will draw the obvious conclusion: the tool changed nothing except the amount of work they are expected to absorb. Leaders who are serious name the thing that stops: the form nobody reads, the summary that is now written once instead of three times, the meeting that shortens. This is the point at which AI adoption stops being a training problem and becomes a leadership one.
4. A sentence about what we will not use AI for, said without checking
Ask a leader what their organization will not use AI for and time the pause. A leader who has led the decision answers immediately and in their own words. A leader who has adopted a policy goes to find the policy. England's Department for Education makes this a practical matter rather than a philosophical one. Its Generative AI: product safety standards, formerly titled product safety expectations, published January 22, 2025 and last updated January 19, 2026, are written for edtech developers and suppliers, and say schools and colleges may also find them helpful in assessing which AI products are safe for use in education. The standards give a school something to measure a product against. They do not say no for you. Somebody in the building has to be able to say no to a product, out loud, and be backed. I have set out how to draw that line so it survives contact with a real case in which decisions AI should never make.
5. A question the leader says out loud they cannot answer
The most reliable marker I know of good AI leadership is a leader who names, in front of staff, the thing they have not worked out yet: what happens to coursework, whether the assessment policy holds, how they will know if any of this improved anything. A leader who admits an open question is a leader who has been thinking about it. A leader with a confident answer to everything has usually been reading vendor material. Openly held questions also do something practical: they tell staff which problems are worth bringing upward, and they make it safe to disagree in a meeting.
What the research says leaders are responsible for, and what it leaves out
Academic work has started to define the AI job of a school leader, which is progress, though it defines responsibilities rather than behaviors. In February 2026, David DeMatthews, Pedro Reyes, Torri D. Hart and Lebon James III published a six-domain framework for AI leadership in schools in Educational Management Administration & Leadership, aligned to the Professional Standards for Educational Leaders: ethical AI use, strategic planning and visioning, curriculum and instructional innovation, professional development and capacity building, data-informed decision-making, and operational efficiency. Their argument, which I agree with, is that AI is now a leadership imperative rather than a technology project.
Here is the limit worth naming. It is a conceptual framework, not a study of what distinguishes schools where AI is led well from schools where it is not, and a leader can hold all six responsibilities and still be invisible on all five of the signs above. Domains tell you what sits on your desk. The Decision Trail tells you whether anything left it. The same instinct shows up in district practice. Reporting the ILO Group's case for adaptive governance in Government Technology on March 30, 2026, Julia Gilban-Cohen quotes Delaware's education secretary, Cindy Marten, relaying Governor Matt Meyer's test for AI: "If it doesn't make learning stronger, safer or fairer, it's just noise." The same article describes Charlotte-Mecklenburg Schools gathering feedback from roughly 10,000 students, families and staff before deploying any AI tools.
The Mistake I See Most Often
The most common mistake is a leadership team that has confused being informed about AI with leading on it. I see it in the newsletter replies most weeks, and in the room when a senior team has read widely, attended the webinars, brought in a speaker, and has still not made a decision that anyone below them could name.
Being informed is comfortable because it has no cost and no week after. Leading has both. The team that reads less and decides more will be further ahead by Christmas, because every dated decision teaches the organization something about how AI questions get settled here, and reading teaches it nothing. If you want the sharper version of this, the ten questions every leadership team should be asking about AI are the ones that cannot be outsourced to a report.
Are you leading AI or managing it? A five-question check
You can test yourself against the five signs in about two minutes by answering five questions with a yes or a no, and by asking a colleague to answer the same five about you. This is my suggested self-check rather than a validated instrument, and it is deliberately blunt.
- Dated decision. Can you name an AI decision from the last 30 days, with a date and a person against it?
- Visible use. Have staff watched you use AI on your own work, including something it got wrong?
- Something stopped. Can you name one thing your organization has stopped doing because of AI?
- A stated line. Can you say what you will not use AI for, without opening the policy?
- An open question. Have you told staff, out loud, one question about AI you cannot yet answer?
Four or five yes answers, confirmed by your colleague, means AI is being led. Two or three means it is being managed: the work is happening, but the leader is not yet visible in it. One or none means it is being described, and the Decision Trail above is the place to start. Where your answer and your colleague's differ, the sign exists in your head and not yet in the building, which is the difference this article is about.
Where to start this term
Fill in the Decision Trail template before your next leadership meeting, using the last five decisions from memory. If you cannot reach five, you have found the work, and the fix is not a strategy day. Make one dated decision this fortnight, attach a name and a subtraction to it, and tell people what changed. Then check it against the wider picture using the seven dimensions of AI readiness, which will show you where the next decision needs to come from.
If your leadership team is working through these questions, this is the kind of work I do through AI strategy sessions and leadership mentoring with senior teams. I write each week about what leadership looks like from the outside in the newsletter.
Sources and further reading
Sources last checked October 13, 2026.
- DeMatthews, D., Reyes, P., Hart, T. D. and James, L., III, "Leadership for artificial intelligence use in schools: A six-domain framework for ethical, equitable, and effective integration", Educational Management Administration & Leadership, February 5, 2026.
- Jobs for the Future, "AI Usage in Education is Growing, But Gaps in Guidance Persist, New Survey Finds", March 18, 2026, reporting an AudienceNet survey of 3,020 learners fielded November 28 to December 8, 2025.
- Department for Education, "Generative AI: product safety standards" (formerly product safety expectations), published January 22, 2025, last updated January 19, 2026.
- SWGfL, "Keeping Children Safe in Education 2026: what do schools need to know?", July 9, 2026, on KCSIE 2026 coming into force on September 1, 2026, the nudes and semi-nudes definition and the annual filtering and monitoring review.
- EdTech Innovation Hub, "KCSIE 2026 published with new AI, mobile phone, and vetting rules for schools", July 7, 2026, on Keeping Children Safe in Education 2026, in force September 1, 2026.
- Gilban-Cohen, J., "School Districts Prioritize AI Governance, Not Adoption Speed", Government Technology, March 30, 2026.
Dan Fitzpatrick is The AI Educator: a Forbes contributor, international keynote speaker and author who works with school and organizational leaders on AI strategy, readiness and leadership. More about Dan.
